docs(memory): document opt-in reality across the wiki
Bring the wiki onto the opt-in memory (commit 35bbec1: Cargo feature `memory`
default off + runtime flag `serve --memory`). Memory was missing or read as
always-on on seven pages.
- Memory.md: new "Enabling memory" section (two gates, paths, rustc 1.89 floor);
corrected "embedded in serve" to opt-in; subtitle notes off-by-default.
- Installation.md: prerequisites rustc floor (lean 1.85+, memory 1.89+),
`--features memory` build command, new "Memory (optional)" setup subsection.
- Usage.md: serve section gains the `--memory` enabling path, /memory/* curl
examples (remember/recall/projects), and the 503-without-flag note.
- Hardware-and-Compatibility.md: "Memory (optional) — CPU, not VRAM" (AVX-512/
VNNI fast path, MLAS fallback; no VRAM, doesn't touch the 16 GB budget).
- Configuration.md: "Memory (optional, default off)" flag table
(--memory / VULKANFORGE_MEMORY / VF_MEMORY_DB).
- Troubleshooting.md: memory FAQs (503, rebuild message, first-start download,
where it's stored).
- Home.md: v1.0 memory section reframed opt-in (two gates, cost only with feature).
Facts verified against the binary (serve --help, default paths) and declared
MSRVs (sqlitegraph rust-version 1.89, ort 1.88), not the brief. No Stufe-B docs
(REPL/agent memory tools don't exist yet). Goes out with the code release (mg).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
docs(wiki): v0.8.0 — auto ctx-size, Gemma-4 tool-calling, vf-clide; KV-FP8 now required
- New page: vf-clide (the standalone CLI chat client) + Sidebar/Home/Usage/Installation links.
- Automatic context sizing (v0.8.0): new Configuration section, Usage/Home/Troubleshooting notes,
the 16384 RDNA4 LDS ceiling (explicit --ctx-size above it aborts, not clamped).
- Gemma-4 native tool/function calling documented (Usage, Supported-Models).
- KV-FP8 corrected from "recommended" to REQUIRED for the Gemma-4-26B-A4B MoE across
Supported-Models / Configuration / Usage / Troubleshooting / Architecture / Choosing-a-Model:
the non-FP8 KV path is known-broken and the engine fail-loud aborts without VULKANFORGE_KV_FP8=1
(debug override VULKANFORGE_ALLOW_BROKEN_KV=1).
- Home/Configuration now reference the shipped v0.8.0 (perf matrix provenance stays v0.7.0).
docs(wiki): add VF_USE_GRAPH to Configuration
Opt-in graph-dispatch flag (default off; topologically-sorted per-layer dispatch vs the default
imperative loop). Effect + default sourced from the v0.7.0 code; neutral wording.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
docs(wiki): VulkanForge user wiki @ v0.7.0
9 content pages + sidebar + footer, sourced strictly from the v0.7.0-tagged repo (README,
CHANGELOG, docs/BENCHMARKS.md, docs/MODELS.md, Cargo.toml, LICENSE, real flag/CLI code):
Home, Installation, Hardware-and-Compatibility, Supported-Models, Usage, Configuration,
Benchmarks, Architecture, Troubleshooting.
User-facing, English, honest (v0.7.0 same-run matrix vs llama.cpp Vulkan: dense prefill parity
0.93-1.04x, Gemma-MoE @2048 0.83-0.89x, decode 0.87-0.97x). No pre-v0.7 numbers, no internal
(sprint reports / coding-standards / bisect) leaks; flags grepped from code with verified defaults.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>